1. Field of the Invention
The invention relates generally to switches used in communications networks. In particular, it relates to optical packet-based communication networks.
2. Background Art
An increasing fraction of switched long-distance telecommunications is being performed based on the Internet Protocol (IP). A conventionally conceived Internet-type of communication network 10, as schematically illustrated in FIG. 1, connects multiple terminals 12 through nodes 14 interconnected by bi-directional communications links 16. The terminals 12 can be considered to be ports to other, perhaps different, types of computer networks. The nodes 14 are based on routers which can route sequentially received packets in different directions as the packet propagate through the network 10 from the source terminal 12 to the destination terminal 12. The node routers switch the packets in different directions dependent upon address labels contained in the individual packets.
While IP was originally designed to work with any physical layer transport medium, it was developed for data networks based on electrical cables of different sorts and capacity. Electronic routers are readily and economically available that operate at the speeds of electrical cable, typically about 155 megabits per second (Mb/s) and below.
However, in the same time period in which IP was being extensively implemented, long-distance telecommunication networks have become increasingly based upon optical fiber as the transmission medium. Fiber allows the data rates in a channel to increase to 1 gigabit per second (Gb/s) and even 10 Gb/s, the speeds being limited by the electronics and opto-electronics at the transmitter and receiver. Further increases in speed will be difficult. Nonetheless, electronic routers are available for these data rates.
The total data rate of an optical fiber can be significantly increased by wavelength division multiplexing (WDM) in which a single fiber conveys multiple optical carriers of number W of different wavelength, each impressed with its own data signal. Multiple sets of electronics and opto-electronics operating in parallel at the transmitter and receiver respectively generate and detect the respective optical signals. At the transmitter, the data signal modulates a laser outputting at a selected WDM wavelength, and the different modulated carriers are combined onto the single fiber. At the receiver, an optical detector receives a wavelength-separated WDM channel and converts its envelope to electrical form. Optical means, such as a diffraction grating or an array waveguide grating, can combine or separate the optical signals at the ends of the fiber. Even if the electronics are limited to separate data rates in the low gigahertz range, the fiber throughput or total data rate is W times greater. Some of the earlier WDM systems carried only 4 separate channels, but more advanced systems, referred to as dense WDM (DWDM), have been proposed in which 80 and more WDM channels are impressed on the fiber, thus vastly increasing the total data rate.
Although it is possible to use electronic routers with WDM or DWDM, they do not scale well with increasing number of WDM wavelengths. A router operating with W=32 WDM wavelengths and having K=4 multi-wavelength input ports and K=4 multi-wavelength output ports and operating at 10 Gb/s has a total aggregate switching capacity of 1.28 terabits per second (Tb/s). However, a non-blocking electronic switch needs to connect any input port to any output port regardless of color. Such switching fabric is available in a Clos network, but the power and complexity increases as (KW)2 and 4·21/2(KW)1.5 respectively. If presently available electronic routers are scaled to 10 Tb/s capacity, the router requires 54 bays of electronics weighing 400 kg and consuming 400 kW of power, clearly an uneconomical design.
In U.S. patent application Ser. No. 10/081,396, filed Feb. 22, 2002 and incorporated herein by reference in its entirety, Yoo describes an optical router relying on wavelength conversion that routes individual packets without converting their payload to electrical form. Its switching fabric 20 is schematically illustrated in FIG. 2. The outputs of K input fibers 22 each carrying W wavelength-separated channels are connected to respective demultiplexers 24 which separate the W WDM signals to be input to respective tunable input wavelength converters 26. Each wavelength converter 26 can be dynamically tuned to convert an input signal at one wavelength to any one of a plurality of wavelengths, preferably of number WK. The wavelength conversion leaves intact any data impressed upon the optical signal. That is, the optical carrier wavelength is changed while the carrier modulation is left intact. A WK×WK wavelength router 28 receives the outputs of the input wavelength converters 26 on its input ports and routes them to any one of its output ports according to its wavelength. That is, the routing is determined by tuning the input wavelength converters 26, and in the described embodiment the wavelength at a router output port is fixed so that tuning the wavelength at the input port to that of the desired output port accordingly routes the signal. The wavelength router 28 can be implemented in an array waveguide grating (AWG), which is a passive optical waveguide structure capable of combining or separating a multi-wavelength optical signal. A WK×WK AWG provides non-blocking switching, but such an AWG having large values of WK is difficult to fabricate. As described by Yoo, however, a number M of smaller AWGs allows a reduction in the number of routing wavelengths by M2, and the AWGs can operate in parallel with different sets of the WDM wavelengths to provide only limited selection of wavelength, and such a design does not markedly increase the blocking at reasonable traffic loads.
The output ports of the wavelength router 28 are connected to the inputs of respective output wavelength converters 30 which convert the optical carrier wavelength associated with that output port to a selected one of the W WDM wavelengths, again without disturbing the data modulation of the carrier. Because both input and output wavelengths at the output wavelength converter 30 are fixed for a particular port, the output wavelength conversion need not be tunable. The tunable and fixed wavelength converters 26, 30 may have the same basic structure, for example, a tunable or untuned laser and a Mach-Zehnder interferometer receiving both the laser radiation and the unconverted but modulated optical signal and outputting an optical signal with the laser wavelength but with the same modulation as the input signal. Optical multiplexers 32 receive the outputs of W of the output wavelength converters 20 and combine them into a multi-wavelength optical signal coupled to an output fiber 34. Both the demultiplexers 24 and multiplexers 32 may also be AWGs although other structures are possible.
For increased traffic capacity, the input-wavelength converters 26 should be tunable on time scales on the order of the duration of an IP packet at the intended high data rates, for example, in less than 10 ns. An optical IP packet 40 illustrated in the timing diagram of FIG. 3 includes a payload 42 and an optical header 44. The IP payload 42 is of variable length, for example, up to about 1500 bytes. The IP header 44 on the other hand is typically of fixed length. An IP datagram header 46, as illustrated in FIG. 4, typically has a fixed length of 160 bits (20 bytes) including, among other things, 32-bit source and destination addresses and a datagram length. The source and destination addresses at some level are uniquely identified to each of the users 12. The router includes a lookup table that converts destination addresses to an output path from that router. Both the payload 42 and the optical header 44 propagate together across the IP network. That is, contents of the header 44 determine the switching at each node, which is performed according to the tuning of the input wavelength converters 26. For this reason, the described switching is referred to as optical label switching (OLS). With a few exceptions such as time-to-live counters, the header 44 remains unchanged as it propagates and is switched at intermediate routers together with the associated payload. There are several choices of formatting the optical header 44 relative to the payload 42. The header 44 may precede the payload 42 in a single serial data stream, as in done in conventional electrical signaling. Alternatively, the header 44 may be impressed on a separate WDM wavelength channel used for all the headers propagating from one node to the next so that they accompany the payloads 42 of different wavelengths. Most preferably, as will be explained later for the invention, the header 44 is impressed on a separated sub-carrier modulation channel on the same WDM wavelength channel as the data payload 42 so that each header and payload pair will be transported on the same carrier wavelength.
Such an optical switch is capable of routing a large IP traffic. However, there are some problems in integrating it into a complete IP network. Such an optical router is needed for the high-capacity core of an IP network, for example, the long-distance portion of the network, typically a public network, which interconnects smaller, more local networks. The local networks can continue to use conventional electronic routers. Networks within buildings typically use electrical cables rather than optical fibers, and even point-to-point optical fiber links are easily accommodated with conventional electronic routers. However, we have observed that the optical label switching does not mesh well with electronic routing.
A router must switch IP packets arriving asynchronously from a number of input ports to selected ones of its output ports. Contention arises if two packets arriving from different input links require switching to the same output link. To overcome contention, a router includes a buffer or queue, for example, a first-in/first-out (FIFO) register, which should be large enough to store enough packets for sufficient time that the contention can be resolved. IP networks are stochastic and non-deterministic so that some periods of very high traffic will occur. It is expected that a certain number of packets will be lost because of buffer overflow or other reasons. The system is designed to overcome some loss, but excessive loss becomes unacceptable. Excessive loss is usually associated with congestion in which a node experiences too much contention and the buffers overflow.
An associated problem arises because there may be different attributes of service for different packets, as indicated by an 8-bit field in the header in this example. Priorities may be set according to various and different criteria, for example, a quality of service (QoS) requiring no more than a predetermined maximum loss of data, type of service (ToS), and class of service (CoS). Some time critical applications require no more than a predetermined maximum delay for propagation across the network, also called latency. As described later for the invention, there may be separate output buffers assigned to packets addressed for one destination and having different ones of these attributes. It is understood that a single random-access memory (RAM) may be used for all buffers with pointers to the RAM controlling the queuing.
Electronic buffers are easily implemented in RAM and can be made relatively large. Optical buffers, on the other hand, are not so readily available. The technology of optical RAM has not been sufficiently developed for use in routers. Heretofore, optical queuing has been accomplished in most part by fiber delay lines providing a fixed delay before the payload is inserted into the switch fabric. A 50 m length of fiber introduces about 250 ns of delay, which may be sufficient time to resolve the contention. The delay may be introduced at the input to the router after the header information has been extracted. If necessary, one or more fiber delay lines may link pairs of input and output ports of the wavelength router so that a packet needing additional delay can be looped back from the router output to its input with additional delay introduced. Further, in the wavelength routing described by Yoo, contention may be substantially reduced by the ability to select between multiple wavelength channels linking neighboring nodes.
Nonetheless, it is fair to say that optical label switching of IP packets is more prone to congestion than is electronic switching, primarily because of the limited depth of delay buffers and because of the need to asynchronously accommodate variable length packets without segmenting them as is possible with electronic routers. It is well known that the typical IP traffic on client networks has a distribution of packet lengths that is strongly peaked near 40, 574, and 1500 bytes with almost 50% of the packets having lengths between 40 and 52 bytes. Electronic routers can be fabricated with random access memories easily accommodating this distribution. Nonetheless, when such traffic is combined into a higher-capacity optical core network, the traffic has similar irregular distribution of packet sizes since the auto-correlation function continues to be large for smaller packets even for high traffic. Statistically significant occurrences over short time periods of multiple small packets, that is, high auto-correlation over these periods, necessitate increased buffering. We have determined that such a distribution suffers substantial congestion at optical switching nodes having limited buffering when the optical transmitter load exceeds about 40% of theoretical capacity.
Accordingly, it is desired configure a combination of electronic and optical routing system that does not suffer such incompatibilities.
Another aspect of optical routers utilizing optical label switching is they they provide more than 1000 times greater capacity and speed than conventional electronic IP routers. Accordingly, it is desired to configure a combination of optical and electronic routers that does not suffer such incompatibilities.
Optical routers are expected to be concentrated in the high-capacity main portion of the Internet backbone. It is desired that this portion of the network be able to accommodate the older, non-IP types of traffic so that separate parallel long-distance networks do not need be built or older local networks do not need to be converted to IP. Examples of more traditional formats include, for example, SONET, which has slotted traffic with successive packets in a fixed length and repetitive frame dedicated to a virtual circuit between clients, and optically switched WDM. IP traffic is often accommodated in SONET networks by placing IP packets in dedicated ones of the repeating SONET time slots. It is desired to easily transfer this source of IP traffic onto an IP network with optical routers.
An associated problem involves IP or other networks which have been provisioned with equipment operating at significantly different data rates. It is desired to provide a standard interface from all these networks to the high-speed core network without needing to upgrade legacy networks to the current standards.